Accelerated volume rendering

Abstract

Performing accelerated volume rendering of a scene in a computer system by: loading volumetric data into a first part of a system memory, the volumetric data represented as a three-dimensional array of data values; determining intersection points of each of a set of sample rays cast through the scene and intersecting the three-dimensional array, wherein the intersection points are determined as a function of sample times from beginning to end; using the intersection points, computing a subset of the three-dimensional array for placement into local storage; loading the subset of three-dimensional array into the local storage; and repeating the determining, computing, and loading acts for all sample times from beginning to end for all sets of sample rays cast by the processor device.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS[0001]None.STATEMENT REGARDING FEDERALLY SPONSORED-RESEARCH OR DEVELOPMENT[0002]None.INCORPORATION BY REFERENCE OF MATERIAL SUBMITTED ON A COMPACT DISC[0003]None.FIELD OF THE INVENTION[0004]The invention disclosed broadly relates to the field of volume rendering on a cell broadband engine or more generally relates to accelerated volume rendering on cache-less systems.BACKGROUND OF THE INVENTION[0005]Volume rendering is a visualization technique that renders volumetric data (three-dimensional data). A voxel is an element of volumetric data. The term is a concatenation of the first syllables of “volume” and “picture element,” analogous to the two-dimensional term pixel made from “picture” and “element.” The volumetric data can be acquired from medical scanners, seismic acoustic collection devices, or data generated by algorithmic methods on a computer system. Volume rendering enables the visualization of sampled functions of three spatial dimensi...

AI Technical Summary

Benefits of technology

The patent describes a method to quickly render a computer system's view of a scene by first loading the data that describes the scene into memory. The processor then uses those data to determine specific points where different rays of light intersect with the scene, and creates a smaller set of data that can be placed into a local storage. This process is repeated for all the different rays of light that are cast through the scene. Once all the data is loaded, the system can quickly use the data to create a visual representation of the scene. This method helps to improve the speed and performance of the process of rendering computer system images.

Problems solved by technology

Most ray tracing algorithms are implemented on systems with preemptive multi-tasking, virtual memory management, and multi-level cache, because as the rays traverse the data they typically jump to non-contiguous memory locations resulting in a lot of cache misses.

Both methods result in various performance impacts such as cache misses, page faults, and task switching.

A serious drawback is that only one surface at a time can be extracted.

For example, if one wishes to highlight a blood vessel or other tissue, one must first revert to the original data and extract an entirely different surface—this is an expensive process.

Even so, direct methods do allow path navigation along centerlines, even though there is no explicit surface with which to detect a collision.

Direct methods are predominant today and among direct methods, ray casting is by far the most popular and advantageous, but also the most computationally intensive.

This is an expensive operation because the time latency of moving data from system memory to L3, L2, and L1 cache can be many clock cycles.

During this time, the microprocessor arithmetic unit can sit idle, wasting computer resources and degrading application performance.

The problem is that if data is not in register memory it must be fetched from cache.

This multi-level memory hierarchy makes it difficult to control performance as tracing a ray from the eye through the volume results in many accesses to non-contiguous locations in memory and hence increased probability that data will not be in one of the aforementioned memory locations.

The advantage is predictability in response and overall performance, but the disadvantage is that more programming is required to explicitly manage data movement.

In the case of GPUs (graphics processing unit), DSPs (digital signal processors), or Cell / BE there is no hardware support for this.

To further complicate programmability, memory close to the computation units is usually very small, i.e. too small to store an entire volume of data.

This will definitely result in many small accesses to non-contiguous memory locations and the overhead of these accesses, in the form of cache miss latency, will likely impact overall volume rendering performance.

Additionally, the computational units will be idle for a greater percentage of the time as they wait for the next small piece of data, therefore processor utilization will be lower than ideal.

(1) Some portion of fast access memory must be reserved to cache data.

(2) Some portion of fast access memory must be used to store code that manages this software cache.

(3) Access to data is not immediate, i.e. a caching scheme and the logic that supports it must be invoked. Ultimately this will lead to performance challenges, e.g. if data is not in software cache, a cache miss will be issued and additional logic is used to go fetch the data from the next level of memory in the hierarchy.

Images